Pavel Simakov - JJTree Tutorial for Advanced Java Parsing

 

 

 

 

by Pavel Simakov 2008-06-11

The Problem

JJTree is a part of JavaCC is a parser/scanner generator for Java. JJTree is a preprocessor for JavaCC that inserts parse tree building actions at various places in the JavaCC source. To follow along you need to understand the core concepts of parsing. Also review basic JJTree documentation and samples provided in JavaCC distribution (version 4.0).

JJTree is magically powerful, but it is as complex. We used it quite successfully at my startup www.moola.com. After some the basic research into the grammar rules, lookaheads, node annotations and prototyping I felt quite comfortable with the tool. However, just recently when I had to use JJTree again I hit the same steep learning curve as if I have never seen JJTree before.

How to write a tutorial that gets you back in shape quickly without forcing the full relearning?

The Solution

Here I capture my notes in a specific form that I do not have to face that same learning curve again in the future. You can think my approach as layered improvement to a grammar that follows these steps:

  • get lexer
  • complete grammar
  • optimize produced AST
  • define custom node
  • define actions
  • write evaluator

I always start simple and need to go more complex - this is exactly how I will document it. In each example I start with a trivial portion of grammar and then add some more to it to force specific behavior. New code is always in green. Let's hope this save all of us the relearning.

Reorder tokens from more specific to less specific

The token in TOKEN section can be declared in any order. But you have to pay very close attention to the order because the matching of tokens starts from the top and down the list until first matching token is found. For example notice how "interface" or "exception" are defined before STRING_LITERAL. If we had defined "interface" after STRING_LITERAL "interface" would never get matched,  STRING_LITERAL would. 


TOKEN : {

	  <INTERFACE: "interface" >

	| < EXCEPTION: "exception" >

	| < ENUM: "enum" >

	| < STRUCT: "struct" >

	

	| < STRING_LITERAL: "'" (~["'","\n","\r"])* "'" >

	| < TERM: <LETTER> (<LETTER>|<DIGIT>)* >

	

	| < NUMBER: <INTEGER> | <FLOAT> > 

	| < INTEGER: ["0"-"9"] (["0"-"9"])* >

	| < FLOAT: (["0"-"9"])+ "." (["0"-"9"])* >

	| < DIGIT: ["0"-"9"] >

	| < LETTER: ["_","a"-"z","A"-"Z"] >

} 

The ordering is the same reason why we can't just use "interface" inline in the definition of productions. The STRING_LITERAL will always match first.

Remove some nodes from final AST

Some nodes do not have any special meaning and should be excluded from the final AST.  This is done by using #void like this:


void InterfaceDecl() #void : {

}{ 

	ExceptionClause()

	| 

	EnumClause()

	|

	StructClause()

	|

	MethodDecl()

}

Add action to a production

You will definitely need to add actions to the production for your parser to be useful. Here I capture the text of the current token (t.image) and put it into jjThis node that will resolve to my custom node class TypeDecl. You bind a variable "t" to a token using "="; the action itself is in curly braces right after the production and can refer to current token as "t" and current AST node as "jjtThis".


void TypeDecl() : {

	Token t;

}

{

	<VOID>

	|

	t=<TERM> { jjtThis.name = t.image; } ("[]")?}

}

Here I further set isArray property to true only if "[]" is found after the <TERM>:


void TypeDecl() : {

	Token t;

}

{

	<VOID>

	|

	t=<TERM> { jjtThis.name = t.image; } ("[]" { jjtThis.isArray = true; } )?}

}

Multiple actions inside one production rule

Just as we have seen earlier you can access values of multiple token in one production rule. Notice how I declare two separate tokens "t" and "n". Here:


void ConstDecl() : {

	Token t;

	Token n;

}

{

	LOOKAHEAD(2)

	

	t=<TERM> { jjtThis.name = t.image; } "=" n=<NUMBER> { jjtThis.value = Integer.valueOf(n.image); }

	|

	<TERM>

}

Lookaheads

There are certain points in complex  grammars that might not get parsed unambiguously using just one token look ahead. If you are writing high performance parser you might need to rewrite grammar. But if do not care about performance you can force lookahead for more that one symbol.

JJTree generator will give you a warning about ambiguities. Go the the rule it refers to and set lookahead of 2 or more like this:


void EnumDeclItem() : {}

{

	LOOKAHEAD(2)

	

	<TERM> "=" <NUMBER>

	|

	<TERM>

}

Node return values

It is possible to return nodes from the productions, just like function return values. Here I am declaring the ASTTypeDecl will be returned.


ASTTypeDecl TypeDecl() : {

	Token t;

}

{

	<VOID>

	|

	t=<TERM> { jjtThis.name = t.image; } ("[]" { jjtThis.isArray = true; } )?}



	{ return jjtThis; }

}

Once you start having a lot of expressions in one production it is better to group them together so return statement applies to all of them. The above example will actually result in a bug due to a fact that the return statement is attached to one branch of "|" production and not to both branches. We can easily fix the issue using parenthesis to force order of precendence:


ASTTypeDecl TypeDecl() : {

	Token t;

}

{

	(

		<VOID>

		|

		t=<TERM> { jjtThis.name = t.image; } ("[]" { jjtThis.isArray = true; } )?}

	)



	{ return jjtThis; }

}

Build abstract syntax tree as you go

After you have all production return values you can build AST tree on the fly while parsing. Just provide found overloaded add() methods in the ASTInterfaceDecl class and call them like this:


void InterfaceDecl() #void : {

	ASTExceptionClause ex;

	ASTEnumClause en;

	ASTStructClause st;

	ASTMethodDecl me;

}

	ex=ExceptionClause() { jjtThis.add(ex); }

	| 

	en=EnumClause() { jjtThis.add(en); }

	|

	st=StructClause() { jjtThis.add(st); }

	|

	me=MethodDecl() { jjtThis.add(me); }

}

Use <EOF>

Quite often you can get your grammar written and start celebration when you notice that part of the file is not being parsed... This happens because you did not tell the parser to read all content till the end of file and it feels free to stop parsing at will. Force parsing to reach end of file by demanding <EOF> token at the top most production:


void InterfaceDecl() #void : {

}{ 

	ExceptionClause()

	| 

	EnumClause()

	|

	StructClause()

	|

	MethodDecl()

	|

	<EOF>

}

The Final Word

JJTree works incredibly well. No excuse to regex parsing no more... Don't even try to convince me!

Drop me a line if you need help with JJTree - will be glad to share the experiences with you.

References

  1. The JavaCC FAQ by Theodore S. Norvell